Terminal-Bench 2.1 - Hard
Agents · 2026-05-06
The Hard split of Terminal-Bench 2.1 reports the percentage of hard terminal-use tasks completed successfully; higher scores are better.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 84.4 |
| GPT-5.6 Sol | 84.4 |
| Kimi K3 | 74.4 |
| Claude Fable 5 | 70.0 |
| GPT-5.6 Luna | 70.0 |
| GPT-5.5 | 66.7 |
| Gemini 3.6 Flash | 64.4 |
| Opus 4.8 | 63.3 |
| GPT-5.6 Terra | 63.3 |
| Gemini 3.1 Pro Preview | 60.0 |
| Gemini 3.5 Flash | 60.0 |
| Sonnet 5 | 57.8 |
| Grok 4.5 | 56.7 |
| Muse Spark 1.1 | 52.2 |
| GLM-5.2 | 51.1 |
| Kimi K2.7 Code | 45.6 |
| Qwen3.7-Max | 44.4 |
| Opus 4.7 | 40.0 |
| Composer 2.5 | 40.0 |
| Gemini 3 Flash Preview | 40.0 |
| MiMo-V2.5-Pro | 40.0 |
| GLM-5.1 | 37.8 |
| MiMo-V2.5 | 36.7 |
| Sonnet 4.6 | 34.4 |
| GPT-5.4 Mini | 34.4 |
| DeepSeek-V4-Pro | 31.1 |
| Gemini 3.5 Flash-Lite | 31.1 |
| Qwen3.6 Plus (2026-04-02) | 31.1 |
| MiniMax M3 | 30.0 |
| GPT-5.4 Nano | 25.6 |
| Kimi K2.6 | 25.6 |
| Nemotron 3 Ultra 550B A55B | 24.4 |
| Inkling | 23.3 |
| Qwen3.7-Plus | 23.3 |
| Haiku 4.5 | 22.2 |
| MiniMax M2.7 | 18.9 |
| Grok 4.20 | 16.7 |
| Gemini 3.1 Flash-Lite Preview | 15.6 |
| Mistral Medium 3.5 | 15.6 |
| Grok 4.3 | 14.4 |